kimik3.io/Pricing
Kimi K3 API pricing
The official rates, stamped with the date we read them — and the part the rate card doesn't tell you: what K3's always-on reasoning does to your output bill.
Kimi K3 costs $3.00 per 1M input tokens on a cache miss, $0.30 per 1M on a cache hit, and $15.00 per 1M output tokens. The context window is 1,048,576 tokens. The number that surprises people: K3's reasoning bills at the output rate, and we measured it at 43–77% of total output tokens — so the effective cost of an answer is several times the length of the answer itself.
Rates read from the official pricing page on 2026-07-16. Overhead figures measured by us — how.
The official rates
| Item | Rate | Notes |
|---|---|---|
| Input — cache miss | $3.00 / 1M | Every first call pays this. Cold calls never partially cache. |
| Input — cache hit | $0.30 / 1M | 10× cheaper. Granted in 256-token blocks on a repeated prefix. |
| Output | $15.00 / 1M | Includes reasoning tokens — see below. |
| Context window | 1,048,576 | One full call at ~500k input is $1.49 and 52 seconds. |
Default max_completion_tokens | 131,072 | Lowering it can bill you for nothing. |
These are also the rates you pay through EvoLink — no gateway markup on kimi-k3. That makes the choice a narrow one: same model, same price, but one key that also reaches GPT, Claude and Gemini, and no Chinese phone number or separate Moonshot account to set up.
What reasoning actually costs you
K3 has no reasoning off-switch. Every response contains a reasoning_content field, and every token in it bills at the $15.00 output rate — the same as words you actually read. So the interesting number isn't the rate. It's the share of your output bill you never see.
We measured three prompts on 2026-07-16:
| Prompt | Reasoning | Total output | Reasoning share |
|---|---|---|---|
| "What is 2+2? Answer with the number only." | 32 | 48 | 67% |
| "Explain in one paragraph why prompt caching lowers API cost." | 715 | 930 | 77% |
| A two-train word problem, reasoning shown | 220 | 518 | 43% |
Asking K3 what two plus two is costs 32 tokens of deliberation to produce a one-character answer. Two-thirds of that invoice line is thinking.
Counter-intuitively the hardest prompt had the lowest reasoning share (43%). Three data points do not make a curve, and we are not going to pretend otherwise — the useful takeaway is the magnitude, not a formula: budget several times more output tokens than your answer length, on every prompt, including trivial ones.
Working out a real bill
Take an agent with a 50,000-token fixed prefix serving 200 requests a day — 10M input tokens daily. Input alone:
| Scenario | Rate | Per day | Per 30 days |
|---|---|---|---|
| Every request misses | $3.00 / 1M | $30.00 | $900.00 |
| Prefix stays warm | $0.30 / 1M | $3.00 | $90.00 |
| Difference | — | $27.00 | $810.00 |
Arithmetic on the official 2026-07-16 rates, input tokens only. The cheap row is idealised: the first call of each cold period bills at $3.00, and up to 255 tokens of tail never cache. Output bills at $15.00/1M either way — and on this workload, reasoning could easily exceed the entire input bill. Treat these as the ceiling and floor your prompt layout moves you between, not as a quote.
Where the money actually goes
- Output dominates on short prompts. At 5× the input rate, plus reasoning at 43–77% of output, a chatty workload with small prompts is an output-bill problem. Caching won't help you.
- Input dominates on long context. A single ~500k-token call is $1.49 before K3 writes a word. That's where caching earns its keep — a 10× cut on the biggest line item.
- Empty responses still bill. The reasoning-budget trap charges you full output rate for zero characters. This is the only line on your bill that is pure waste.
Pricing FAQ
How much does the Kimi K3 API cost?
$3.00 per 1M input tokens on a cache miss, $0.30 per 1M on a cache hit, and $15.00 per 1M output tokens, per the official pricing page on 2026-07-16.
Are Kimi K3 reasoning tokens billed?
Yes, at the $15.00 per 1M output rate — the same as the answer. We measured reasoning at 43–77% of total output tokens across three prompts, and it bills even when the response contains no content at all.
Is there a free tier for Kimi K3?
Moonshot's own terms are on their platform. Through EvoLink, new accounts start with 10 free credits — enough to run the quickstart from anywhere and confirm your first call end to end.
How do I get the cached input price?
Repeat a long, byte-identical prefix and keep the varying part last. It is automatic — no parameter required. Details and the measured mechanics are on the prompt caching page.
Run the numbers on your own workload
EvoLink carries kimi-k3 on an OpenAI-compatible endpoint — one key reaches GPT, Claude, Gemini, and dozens of the world's mainstream models, so you can price the same prompt across all of them. 10 free credits, sign up from anywhere.